Dear Renzo

Just for clarity I am assuming here that what you are describing is the following scenario: you have some analyses, for example you may be looking at connectivity changes with age across subjects, a second-level regression analyses, and you are running those analyses separately/independently in two cohorts. Let's say for simplicity that your two cohorts have been both analyzed in CONN and are part of the same CONN project, so you would have the following two analyses set up for replication:

Analysis 1: Subject effects: "Cohort1, Age*Cohort1", Between-subjects contrast "[0,1]"

Analysis 2: Subject effects: "Cohort2, Age*Cohort2", Between-subjects contrast "[0,1]"

and what you observe, if I am underestanding correctly, is that using the default options in CONN second-level ROI-to-ROI analyses (standard setings for cluster-based inferences #1: Functional Network Connectivity) you get different orderings/groupings of ROIs in "results 1" and "results 2". That is of course all perfectly fine, the FNC approach performs hierarchical clustering on the raw data trying to first define groups of "similar" connections on to which then run the statistical analyses, and that clustering, depending on sample size, will introduce some additional variablity across replications (to be precise the order of the original ROIs does not matter, and the order of the ROIs within each resulting group also does not matter, but the nature of the groups of ROIs resulting from the clustering procedure is data-driven so it will vary when new subjects or new data is used, and those changes will be particularly notable for small sample sizes). 

In order to avoid that variability the simplest option is to define groups of ROIs apriori. For example, if you are using ROIs defined in the Schaefer atlas, then you could define groups of ROIs using the classical networks. These networks have effectively been also computed using a similar clustering procedure but just on a much larger dataset. To do that in CONN you would simply click on the label "ROIs sorted using hierarchical clustering" and choose there the appropriate option (for example choose "Use predefined order/groups for Schaefer atlas ROIs"). That will fix the grouping and ordering of the ROIs being analyzed and displayed to a fixed apriori order. In addition, you can also save and load in that same menu your own groupings/orderings from prior analyses and then re-use them here in your current analyses.  

Of couse another option is to use other statistical procedures that, unlike FNC, do not depend on the groupings or orders of ROIs. For example, a common choice is to perform mass-univariate second-level analyses across the entire matrix of ROI-to-ROI connections and then correct the statistical results using for multiple comparisons using FDR. To do that in CONN you would simply click on the label "Standard setings for cluster-based inferences #1: Functional Network Connectivity" and choose there the option "Alternative settings for connection based inferences: parametric univariate statistics". You can still change if you want the order of your ROIs being displayed, but the results statistics when using this option will remain exactly the same across all orderings. 

Hope this helps. Also I did not totally understand what you meant by "remove the order" of the regions so please let me know if this still failed to answer some part of your question. 

Alfonso

Dear Alfonso and CONN users,

I have been strugling for months with what seems to be a limitation of CONN (hopefully is just something I have to learn about the toolbox)

If I compare the Roi to Roi results between 2 diferent cohorts, corrected for cluster level, I run into a strange behaviour of the toolbox: I get VERY different results for even slight changes in the order of the regions in the connectome. Also, if I change the order I get different results that are not at all similar to the previous region order. In other words, you can either not see true results in one cohort or see 2 different results in one cohort. Then, only one of the regions order works in the second cohort. What leaves us with the very dubious task to assume which is "the right" one. 

To make things even more complicated, If you remove the order of the regions, unfortunately thats no garanty that you will get the results in the same order, because even the order of the regions when you "remove the order" changes, is not always shown in a fix order.

Unless there's some procedure which sets some order in the Roi to Roi results, this limitations makes it, unfortunately, unreliable for publication.

Please could anybody help?

Best

Renzo